Emil mentioned off the tracker that the diff looks good to him.

I have merged this. Thanks a lot for reporting this! :-)

Results are in. Full before/after translator results:

https://ai.dmi.unibas.ch/_tmp_files/helmert/issue887-v1-issue887-base-issue887-v1-compare.html

Filtered to the cases where the translator results are not identical before/after:

https://ai.dmi.unibas.ch/_tmp_files/helmert/issue887-v1-issue887-base-issue887-v1-compare-filtered.html

In summary, nothing substantial changes. The translator runs out of memory
before/after in the same 46 IPC 2018 instances. There are only 6 tasks where the
before/after results are different (5 in trucks-strips, 1 in citycar-opt14-adl),
and this can be explained by the known nondeterminism of the translator in these
domains.

The differences in runtime are negligible and dominated by noise. The only part
of the code that is changed is the parsing stage, whose runtime is usually
negligible anyway.

So from the effect on the IPC benchmarks, this looks safe to merge.

Anyone want to have a brief look at the pull request before I merge? Apart from
the code for the experiments, it's quite trivial. (Plus 1 line and minus 8
lines, 5 of which are comments.)

Thanks! (It's not such a large experiment, though, only two configurations and
only the translator, which doesn't produce that much output. But of course we do
compute the md5 hashes of output.sas. Forty minutes was just the time passed so
far. Now we are at more than an hour and counting. :-))

Something on the order of 40 minutes for the "parse-again" step is quite common,
I think. Most parsers search for a lot of regular expressions. If you have an
idea on how to speed up parsing, I'd be interested. The code is in lab/parser.py.

The "parse-again" step is only needed if you want to parse the results again. I
guess, it was a left-over from the earlier experiment and could be removed.

> Experiments are running.

The result collection on the grid (issue887-v1-03-parse-again) after running the
experiment seems to take awfully long. It's already taken more than 40 minutes,
and at until a minute ago it was the only process running on the grid, so I
don't think it's an overloaded file system or similar. Perhaps there is
something not working as expected there, for example related to the special parser?

Also, is this step required at all? I just used the "--all" option when
submitting because that's what we used to do in the past.

Experiments are running.

> Jendrik, can you help me set up a translator experiment? I'd be interested
> in whether or not the output remains the same and how the change affects
> translator runtime.

Never mind, I found the experiments for issue862 and can work from there.

I have updated experiments/README to point to that issue for translator experiments.

Moving the "Optional summary", which includes the original bug report, to a
"Change note". Forgot to do that earlier. (Emil, I'm sorry the UI of the tracker
is very misleading there. The "Optional summary" is meant as an internal TODO
list to track the state of work on the issue.)

=========================================================================
Original report by Emil:
=========================================================================

Replacing a goal

(:goal (and (a) (b) (c)))

with the goal 

(:goal (and (a) (and (b) (c))))

causes a translation failure with the error "Condition not normalized". The full
error 
for the 
attached files is as follows:

Parsing...
Parsing: [0.010s CPU, 0.004s wall-clock]
Normalizing task... [0.000s CPU, 0.000s wall-clock]
Instantiating...
Generating Datalog program... Traceback (most recent call last):
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/translate.py", 
line 724, in <module>
    main()
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/translate.py", 
line 695, in main
    sas_task = pddl_to_sas(task)
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/translate.py", 
line 533, in pddl_to_sas
    reachable_action_params) = instantiate.explore(task)
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/instantiate.py", line 75, in explore
    prog = pddl_to_prolog.translate(task)
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/pddl_to_prolog.py", line 165, in translate
    for conditions, effect in normalize.build_exploration_rules(task):
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/normalize.py", 
line 366, in build_exploration_rules
    proxy.build_rules(result)
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/normalize.py", 
line 65, in build_rules
    app_rule_body = condition_to_rule_body(axiom.parameters, self.condition)
  File "/Users/ekeyder/Sandbox/planners/fast-
downward/builds/release32/bin/translate/normalize.py", 
line 388, in condition_to_rule_body
    assert isinstance(part, pddl.Literal), "Condition not normalized: %r" % part
AssertionError: Condition not normalized: <pddl.conditions.Conjunction object at 
0x107afd810>


translate exit code: 30
Driver aborting after translate

pull request here:

https://bitbucket.org/malte/downward/pull-requests/9/issue887/diff

Does someone want to have a look?

Well, it looks like fixing this is a one-line change, but if you look at the
code, there is a large number of related problems that also need fixing, so I'll
try to fix a few of them in one go.

Jendrik, can you help me set up a translator experiment? I'd be interested in
whether or not the output remains the same and how the change affects translator
runtime.

I can confirm the bug. If I run just normalize.py rather than the full
translator, interestingly the goal is replaced by an axiom, but that axiom is
not normalized:

Goal:
  Atom new-axiom@0()

Axioms:
Axiom new-axiom@0()
  Conjunction
    Atom passenger-at(p0, n4)
    Conjunction
      Atom passenger-at(p1, n6)
      Atom passenger-at(p2, n1)

The intended behaviour is that nested conjunctions are flattened, in which case
there would also be no need to replace the goal with an axiom.

Added the elevators domain file.